I have had some success using a class that wraps SSE types as the Scalar type in Eigen matrices. I'm trying to do something similar with another class that wraps the Knights Corner (i.e., Intel Xeon Phi) 64-byte SIMD types. It seems to work for statically sized matrices, but the storage for dynamically sized matrices is not getting aligned properly. A coworker has fixed this problem by modifying Core/util/Memory.h. Given all the occurrences of the literal constant 16 in Memory.h, I doubt that we are overlooking an existing solution to this alignment issue, but I thought I should check here.

Advice appreciated.

In the default branch you can now ask for 64-bytes alignment with -DEIGEN_MAX_ALIGN_BYTES=64